东北大学学报（自然科学版）

东北大学学报（自然科学版） ›› 2021, Vol. 42 ›› Issue (3): 310-317.DOI: 10.12068/j.issn.1005-3026.2021.03.002

• 信息与控制 • 上一篇    下一篇

基于校正分布差异的标定迁移方法研究

赵煜辉， 芦鹏程， 刘晓东， 齐天舒   

  1. (东北大学秦皇岛分校 计算机与通信工程学院， 河北 秦皇岛066000)
  • 收稿日期:2020-09-09 修回日期:2020-09-09 接受日期:2020-09-09 发布日期:2021-03-12
  • 通讯作者: 赵煜辉
  • 作者简介:赵煜辉(1971-)，男，河北秦皇岛人，东北大学秦皇岛分校教授.
  • 基金资助:
    国家自然科学基金青年基金资助项目(61601104).

Research on Calibration Transfer via Correcting Distributions Difference

ZHAO Yu-hui， LU Peng-cheng， LIU Xiao-dong， QI Tian-shu   

  1. School of Computer and Communication Engineering，Northeastern University at Qinhuangdao，Qinhuangdao 066000， China.
  • Received:2020-09-09 Revised:2020-09-09 Accepted:2020-09-09 Published:2021-03-12
  • Contact: ZHAO Yu-hui
  • About author:-
  • Supported by:
    -

RichHTML

27

PDF (PC)

220

摘要： 针对近红外光谱数据的维度高、特征之间存在严重的多重共线性的特点，提出了无迁移标准的通过校正分布差异的标定迁移方法(calibration transfer via correcting distributions difference，CT-CDD).CT-CDD首先建立主仪器的偏最小二乘模型，然后通过偏最小二乘模型提取主仪器和从仪器的潜变量，并且分别对主仪器和从仪器的潜变量进行聚类.该方法基于这样的假设:聚类后的主仪器和从仪器的每一部分特征光谱均服从单高斯分布.最后，找到2个仪器的最接近的子分布，通过校正均值和方差来校正数据分布的差异.实验结果表明CT-CDD通常更加鲁棒并且还可以实现最低的均方根预测误差.

关键词: 近红外光谱;迁移标准;标定迁移;偏最小二乘模型;分布差异

Abstract: Aiming at the characteristics of high dimensionality of near-infrared spectroscopy data and serious multi-collinearity between features， a method of calibration transfer via correcting distributions difference(CT-CDD) was proposed without transfer standards. CT-CDD firstly establishes PLS(partial least square) model of the master instrument， and then latent variables of both the master instrument and the slave instrument are extracted by the PLS model. Next， the latent variables of the two instruments are clustered. The method is based on the assumption that the characteristic spectra of each part of the master and slave instruments after clustering is a single Gaussian distribution. Finally， the nearest sub-distribution of the two instruments is found， and the differences in data distribution is corrected by correcting mean and variance. The experimental results show that CT-CDD is generally more robust and can also achieve the lowest RMSEP (root mean squared error on prediction).

Key words: near-infrared spectroscopy; transfer standard; calibration transfer; partial least square (PLS) model; distributions difference

中图分类号: